本期机器人顶刊精选

Despite the rise of billion-parameter foundation models trained across thousands of graphical processing units (GPUs), similar scaling gains have not been shown for humanoid control. Current neural controllers for humanoids remain modest in size, target a limited set of behaviors, and are trained on a handful of GPUs. We show that scaling model capacity, data, and compute yields a generalist humanoid controller capable of natural, robust whole-body movements. We position motion tracking as a scalable task for humanoid control, leveraging dense supervision from diverse motion-capture data to acquire human motion priors without manual reward engineering. We build a foundation model for motion tracking by scaling along three axes: network size (1.2 to 42 million parameters), dataset volume (more than 100 million frames from 700 hours of motion capture), and compute (21,000 GPU hours). Beyond demonstrating the benefits of scale, we further show downstream utility through a real-time kinematic planner that bridges motion tracking to tasks such as navigation, enabling natural and interactive control, as well as a unified token space that supports virtual reality (VR) teleoperation and vision-language-action (VLA) models with a single policy. Through this interface, we demonstrate autonomous VLA-driven whole-body locomanipulation requiring coordinated hand and foot placement. Scaling motion tracking exhibits favorable properties: Performance improves steadily with compute and data diversity, and learned policies generalize to unseen motions, establishing motion tracking at scale as a practical foundation for humanoid control.

Humanoid robots stand at the forefront of robotics, aiming to capture the agility, robustness, and expressivity of human movement in an anthropomorphic form. The locomotion control of humanoids has evolved from classical model–based methods to reinforcement learning powered by large-scale simulation and now to generative models that produce adaptive, whole-body behaviors, propelling humanoids toward operation in real-world environments. This survey positions humanoid control at a turning point, converging toward a unified paradigm of physics-guided generative intelligence that integrates optimization, learning, and predictive reasoning. We identify three core principles linking these paradigms: physics-based modeling, constrained decision-making, and adaptation to uncertainty. Building on these connections, we provide recommendations for researchers and outline open challenges in safety, accessibility, and human-level capability. These directions represent a transformation from engineered stability to intelligent autonomy, laying the groundwork for humanoid generalists capable of operating safely, collaborating naturally, and extending human capability in the open world.

Humanoid soccer poses a representative challenge for embodied intelligence, requiring robots to coordinate agile locomotion with unreliable visual perception in dynamic environments. However, existing systems typically rely on modular pipelines that separate perception from control or assume ideal sensing, making it difficult to achieve coherent and reactive behavior under real-world perceptual limitations. In this work, we present a unified reinforcement learning–based controller that enables humanoid robots to learn vision-driven reactive soccer skills by directly coupling visual perception with locomotion control. The robot is trained in simulation to acquire soccer behaviors, and adversarial motion priors guide policy learning toward natural motion patterns. To support robust performance under imperfect sensing, we introduce an encoder-decoder architecture together with a virtual perception system that models key characteristics of onboard vision, exposing the policy to perceptual noise and detection failures during training. This design encourages the policy to internalize perceptual uncertainty and continuously adapt its motion in a closed loop. The resulting controller produces coordinated soccer behaviors using only onboard vision, including ball searching, chasing, and multidirectional kicking. It reduces ball position estimation error by 46% and shortens time-to-kick by up to 64% compared with a rule-based baseline, achieving around 90% kicking success in frontfield positions. Experiments across diverse environments and dynamic scenarios, including real RoboCup competitions, further demonstrate the robust performance of the controller. These results highlight the practical effectiveness of integrating perceptual uncertainty directly into policy learning for achieving reliable vision-driven behaviors in humanoid robots operating under real-world conditions.

The Will Smith 2004 blockbuster I, Robot predicts pervasive humanoid robots in 2035; investors agree, roboticists disagree.

As robots enter homes, workplaces, and health care settings, sustaining trust during social interaction becomes a central challenge for human-robot interaction. However, relatively little is known about how humans integrate signals across the brain, hormones, mind, and behavior when robots violate expectations or display social expressiveness. Addressing this gap, we examined how robot performance (congruent versus erroneous) and expressiveness (animated versus stationary) shape multilevel human responses during face-to-face decision-making with an embodied humanoid robot. Participants engaged with the robot in person while neural activity was monitored using functional near-infrared spectroscopy, alongside salivary oxytocin assays, self-reported trust, and behavioral influence measures. Robot errors, implemented as cooperative norm violations, reliably reduced trust and influence, establishing performance reliability as the foundation of trust. Expressiveness amplified these effects: Animated robots elicited stronger prefrontal engagement and cross-level neural-hormonal coupling. Elevated oxytocin was most strongly linked to reduced trust during expressive robot errors, alongside diminished behavioral influence. This pattern is consistent with a context-sensitive vigilance response, indicating that oxytocin in human-robot interaction may heighten sensitivity to norm violations rather than reliably promote bonding. Validation analyses provided small, directionally consistent support for this pattern under counterbalanced order and improved temporal separation. Together, these findings establish a multilevel framework for studying trust in human-robot interaction and reveal a critical design trade-off: Expressive design enhances engagement but can make robot errors disproportionately damaging to trust. These insights identify a biologically grounded boundary condition for oxytocin’s role in social interaction and inform the design of socially effective and trustworthy robots.

Achieving robust, humanlike whole-body control on humanoid robots for agile, contact-rich behaviors remains a central challenge, demanding heavy per-skill engineering and a brittle process of tuning controllers. We introduce ZEST (zero-shot embodied skill transfer), a streamlined motion-imitation framework that trains policies via reinforcement learning from diverse sources—high-fidelity motion capture, noisy monocular video, and non–physics-constrained animation—and deploys them to hardware zero-shot. ZEST generalizes across behaviors and platforms without relying on contact labels, reference or observation windows, state estimators, or extensive reward shaping. Its training pipeline combines adaptive sampling, which focuses training on difficult motion segments, and an automatic curriculum using a model-based assistive wrench, together enabling dynamic, long-horizon maneuvers. We further provide a procedure for selecting joint-level gains from approximate analytical armature values for closed-chain actuators, along with a refined model of actuators. Trained entirely in simulation with moderate domain randomization, ZEST demonstrated broad generality. On Boston Dynamics’ Atlas humanoid, ZEST learned dynamic, multicontact skills (army crawl and breakdancing) from motion capture. It transferred expressive dance and scene-interaction skills, such as box climbing, directly from videos to Atlas and the Unitree G1. Furthermore, it extended across morphologies to the Spot quadruped, enabling acrobatics, such as a continuous backflip, through animation. Together, these results demonstrate robust zero-shot deployment across heterogeneous data sources and embodiments, establishing ZEST as a scalable interface between biological movements and their robotic counterparts.

Accurate measures of muscle forces are critical for engineers, clinicians, and biomechanists, yet direct measurement is highly invasive, and current estimation methods remain limited in accuracy. Here, we demonstrate the application of ultra-wideband radar sensors to noninvasively estimate muscle forces by measuring changes in the electromagnetic properties of contracting muscles, in muscles with different structural properties, during various static and dynamic conditions, and in the presence of fatigue. First, we show that ultra-wideband radar scans of muscle can reliably track isometric force in a unipennate knee extensor (vastus lateralis) and a bipennate ankle dorsiflexor (tibialis anterior). Next, we integrate radar signals in machine learning and linear models to estimate muscle forces during fatiguing isometric and dynamic knee extension contractions, with exceptional accuracy. Last, we identify radar frequency–dependent effects of the relationship between muscle forces and ultra-wideband radar signals, with these effects being independent of physiological and structural features known to influence muscle force. Together, these findings establish ultra-wideband radar as a noninvasive sensor capable of quantifying in vivo muscle forces, with the potential for wearable assistive technologies, biomechanics, and rehabilitation.

Clothes manipulation, such as folding or hanging, is a critical capability for home service robots. Despite recent advances, most existing methods remain limited to specific clothes types and tasks due to the complex, high-dimensional geometry of clothes. This paper presents CLothes mAnipulation with Semantic keyPoints (CLASP). CLASP aims at general-purpose clothes manipulation across diverse clothes types, such as T-shirts, shorts, skirts, long dresses, … and across different tasks, such as folding, flattening, hanging, …. The key idea of CLASP is semantic keypoints —for example, “left sleeve” and “right shoulder”—a sparse spatial-semantic representation, salient for both perception and action. Semantic keypoints of clothes can be reliably extracted from RGB-D images and provide an effective representation for a wide range of clothes manipulation policies. CLASP uses semantic keypoints as an intermediate representation to connect high-level task planning and low-level action execution. At the high level, it exploits a vision-language model (VLM) to predict task plans over the semantic keypoints. At the low level, it executes the plans with the help of a set of pre-built manipulation skills conditioned on the keypoints. Extensive simulation experiments show that CLASP outperforms state-of-the-art baseline methods on multiple tasks across diverse clothes types, demonstrating strong performance and generalization. Further experiments with a Franka dual-arm system on four distinct tasks—folding, flattening, hanging, and placing—confirm CLASP’s performance on real-life clothes manipulation.

Multi-robot systems must simultaneously optimize competing objectives while maintaining coordinated behavior. Existing multi-agent reinforcement learning approaches often rely on fixed or centralized coordination, which limits adaptability and violates distributed constraints. This work introduces the Coordination-Informed Multi-Objective Reinforcement Learning (CIMORL) framework, integrating a distributed weight prediction mechanism, a privileged expert training strategy, and theoretical guarantees for Pareto-optimal solutions. We present the base CIMORL method alongside two sampling-based variants, CIMORL-TS (Tree Search) and CIMORL-MPPI (MPPI), which leverage privileged global information during training to enable fully decentralized deployment. Experimental validation in cooperative and adversarial scenarios demonstrates a $21.2\%$ hypervolume improvement and superior policy stability compared to state-of-the-art baselines. Real-world experiments with Crazyflie drones further validate the framework's robustness in resource allocation and multi-attacker multi-defend scenarios under partial observability.

Autonomous aerial tracking with drones offers vast potential for surveillance, cinematography, and industrial inspection applications. While single-drone tracking systems have been extensively studied, swarm-based target tracking remains underexplored, despite its unique advantages of distributed perception, fault-tolerant redundancy, and multidirectional target coverage. To bridge this gap, we propose a novel decentralized LiDAR-based swarm tracking framework that enables visibility-aware, cooperative target tracking in complex environments, while fully harnessing the unique capabilities of swarm systems. To address visibility, we introduce a novel Spherical Signed Distance Field (SSDF)-based metric for 3-D environmental occlusion representation, coupled with an efficient algorithm that enables real-time onboard SSDF updating. A general Field-of-View (FOV) alignment cost supporting heterogeneous LiDAR configurations is proposed for consistent target observation. Swarm coordination is enhanced through cooperative costs that enforce inter-robot safe clearance, prevent mutual occlusions, and notably facilitate 3-D multidirectional target encirclement via a novel electrostatic-potential-inspired distribution metric. These innovations are integrated into a hierarchical planner, combining a kinodynamic front-end searcher with a spatiotemporal $SE(3)$ back-end optimizer to generate collision-free, visibility-optimized trajectories.Deployed on heterogeneous LiDAR swarms, our fully decentralized implementation features collaborative perception, distributed planning, and dynamic swarm reconfigurability. Validated through rigorous real-world experiments in cluttered outdoor environments, the proposed system demonstrates robust cooperative tracking of agile targets (drones, humans) while achieving superior visibility maintenance.

Language-conditioned robot manipulation is an emerging field aimed at enabling seamless communication and cooperation between humans and robotic agents by teaching robots to comprehend and execute instructions conveyed in natural language. This interdisciplinary area integrates scene understanding, language processing, and policy learning to bridge the gap between human instructions and robot actions. In this comprehensive survey, we systematically explore recent advancements in language-conditioned robot manipulation. We categorize existing methods based on the primary ways language is integrated into the robot system, namely language for state evaluation, language as a policy condition, language for cognitive planning and reasoning, and language in unified vision-language-action models. Specifically, we further analyze state-of-the-art techniques from five axes of action granularity, data and supervision regimes, system cost and latency, environments and evaluations, and task specification. Additionally, we highlight the key debates in the field. Finally, we discuss open challenges and future research directions, focusing on potentially enhancing generalization capabilities and addressing safety issues in language-conditioned robot manipulators.

Collisions of rigid-link robots and rigid environments are often modeled as instantaneous events. Under this idealization, the impact forces become impulsive and the system velocities nonsmooth. In this work, we systematically analyze pre- and post-impact velocities focusing on what we refer to as the nonsmooth impact direction (NSID). We show that it is a characteristic direction of a robotic impact and largely independent of contact properties. The results are directly applicable to large classes of backdrivable robotic systems with rigid links. We address particularities of systems with nonelastic and flexible joints, unconstrained as well as constrained systems. Further, we show that the approach direction w.r.t the NSID sets the direction of the impulsive force in frictional, inelastic impacts. The comprehensive theoretical analysis of this work supported by an experimental validation may serve as a foundation for future planning and control algorithms for various robotic impact applications. These can include humanoid locomotion on a slippery surface or repetitive hammering.

We are entering an era where AI enables multipurpose applications, but the mechanical hardware must be done right.

Real-world robotic manipulation in homes and factories demands reliability, efficiency, and robustness that approach or surpass skilled human operators. We present a real-world reinforcement learning (RL) framework, RL-100, for achieving complete task success under a predefined evaluation protocol built on diffusion visuomotor policies. RL-100 unifies imitation and RL under a single clipped proximal policy optimization surrogate objective applied in the denoising process, yielding conservative, stable improvements across offline and online stages. To meet deployment latency, a lightweight consistency distillation compresses multistep diffusion into a one-step controller for high-frequency control. The framework is task, embodiment, and representation agnostic and supports both single-action and action-chunking control. We evaluated RL-100 on eight diverse real-robot tasks, from pushing and bowling to pouring, cloth folding, unscrewing, multistage juicing, and long-horizon box folding. Under our predefined protocol, RL-100 achieved 100% success in the evaluated trials (1000 of 1000 episodes), including up to 250 of 250 consecutive trials on one task. It matched or surpassed expert teleoperators in time to completion. Without retraining, a single policy attained ∼90% zero-shot success under environmental and dynamics shifts, adapted in a few-shot regime to substantial task variations (86.7%), and remained robust to human perturbations (about 96%). Our juicing robot served customers continuously for about 7 hours without failure when deployed zero-shot in a shopping mall. These results suggest a potential path to deployable robot learning by starting from human priors, aligning training objectives with human-grounded metrics, and reliably extending performance beyond human demonstrations.

Model Predictive Control (MPC) provides interpretable, tunable locomotion controllers grounded in physical models, but its robustness depends on frequent replanning and is limited by model mismatch and real-time computational constraints. Reinforcement Learning (RL), by contrast, can produce highly robust behaviors through stochastic training but often lacks interpretability, suffers from out-of-distribution failures, and requires intensive reward engineering. This work presents a GPU-parallelized residual architecture that tightly integrates MPC and RL by blending their outputs at the torque-control level. We develop a kinodynamic whole-body MPC formulation evaluated across thousands of agents in parallel at 100 Hz for RL training. The residual policy learns to make targeted corrections to the MPC outputs, combining the interpretability and constraint handling of model-based control with the adaptability of RL. The model-based control prior acts as a strong bias, initializing and guiding the policy towards desirable behavior with a simple set of rewards. Compared to standalone MPC or end-to-end RL, our approach achieves higher sample efficiency, converges to greater asymptotic rewards, expands the range of trackable velocity commands, and enables zero-shot adaptation to unseen gaits and uneven terrain.

Seamless operation of mobile robots in challenging environments requires low-latency local motion estimation and accurate global localization. While most sensor-fusion approaches are designed for specific scenarios, this work introduces a flexible open-source solution for task- and setup-agnostic multimodal sensor fusion distinguished by its generality and usability. Holistic Fusion formulates sensor fusion as a combined estimation problem of i) the local and global robot state and ii) a (theoretically unlimited) number of dynamic variables, including automatic alignment of reference frames; this formulation fits countless real-world applications without conceptual modifications, offering a comprehensive solution beyond hard-coded/task-specific approaches. The proposed factor-graph formulation enables direct fusion of an arbitrary number of absolute, local, and landmark measurements expressed with respect to different frames by explicitly including them as states in the optimization and modeling their evolution as random walks. Moreover, local smoothness and consistency receive particular attention to prevent estimation jumps. Holistic Fusion enables low-latency and smooth online state estimation on typical robot hardware while simultaneously providing low-drift global localization at the IMU measurement rate. The efficacy of this released framework [1] is demonstrated in five real-world scenarios on three robotic platforms with distinct task requirements, highlighting the advantages of fusing multiple absolute measurement types [2]. [1] Code: https://github.com/leggedrobotics/holistic_fusion [2] Project: https://leggedrobotics.github.io/holistic_fusion

Humanoid loco-manipulation requires coordinated task-space motion planning with stable loco-manipulation command tracking under complex robot-environment dynamics and long-horizon tasks. While diffusion policies (DPs) show promise for learning from demonstrations, deploying them on humanoids poses critical challenges: the motion planner trained offline is decoupled from the loco-manipulation controller, leading to poor command tracking, compounding distribution shift, and task failures. The common approach of scaling demonstration data is prohibitively expensive for high-dimensional humanoid systems. To address this challenge, we present REFINE-DP (REinforcement learning FINE-tuning of Diffusion Policy), a hierarchical framework that jointly optimizes a DP motion planner and an RL-based loco-manipulation controller. The DP is fine-tuned via a PPO-based diffusion policy gradient to improve task success rate, while the controller is simultaneously updated to accurately track the planner's evolving command distribution, reducing the distributional mismatch that degrades motion quality. We validate REFINE-DP on a humanoid robot performing loco-manipulation tasks, including door traversal and long-horizon object transport. REFINE-DP achieves an over 90% success rate in simulation, even in out-of-distribution cases not seen in the pre-training data, and enables real-world execution without privileged state information. Our proposed method substantially outperforms pre-trained DP baselines and demonstrates that RL fine-tuning is key to reliable humanoid loco-manipulation. https://refine-dp.github.io/REFINE-DP/

LiDAR place recognition supports loop closure, relocalization, and multi-agent map management. As robotic platforms increasingly combine LiDARs with different fields of view, resolutions, and scanning patterns, existing descriptors degrade because they are tightly coupled to sensor-specific characteristics. We present InLiER, a learning-free pipeline based on an intermediate tokenization step. Height-sliced keypoints from structural elements receive mixed-radix token IDs encoding height, radial distance, local shape, and azimuth from local 3D geometry, in a compact sub-2KB representation. The same vocabulary is reorganized across three retrieval stages: height-ceiling histogram intersection for fast rotation-invariant shortlisting, binary bitmask alignment for yaw estimation and reranking, and token-guided geometric verification for 6-DoF pose estimation. InLiER achieves state-of-the-art performance on the HeLiPR dataset and in real-world field experiments, among modern handcrafted methods and outperforms the learning-based baseline on most cross-sensor configurations.

The architecture of apple orchards can be converted to easier operation by implementing the use of a low‐cost, tractor‐mounted, dual manipulator based on a robotic harvesting system with lower degrees of freedom (DoF). To accurately avoid injuries by detecting apples, an efferent, real‐time vision system that includes a new gripping mechanism was developed. In this context, by introducing a new gripping system, spindle wall‐type apple orchard architectures have high potential for easy harvesting using a robotic platform. Therefore, this study was conducted to design and develop a tractor‐mounted, machine vision‐based, 3‐DoF, dual telescopic robotic cartesian system with a novel 2‐DoF gripping mechanism for apple harvesting. The machine vision system was developed using the EfficientDet object detection deep neural network with a mAP@0.5 of 0.775. The tracking algorithm DeepSORT was selected to track the localized apples and transfer the G ‐code to the robotic arm system. The developed novel gripping mechanism consists of three vacuum‐powered, (400 mmHg) spring‐loaded, silicone fingers that provide 0.73 N.m (±0.1) twisting torque to apples in a natural orientation (90°) and a dragging force of 17.65 N (±0.5). The average harvesting cycle time was approximately eleven seconds for both robotic arms, including detection and completion of the harvesting cycle.

The coordination of Automated Guided Vehicles (AGVs) in high-density industrial environments represents a critical challenge within Logistics 4.0, as traditional traffic management methods often lead to inefficiencies caused by negotiation-based priority assignment. To overcome the resulting limitations, this paper presents an innovative AGV traffic management system based on a Lifelong Multi-Agent Path Finding (L-MAPF) algorithm operating on roadmaps generated with Non-Uniform Rational B-Splines (NURBS) curves. The approach guarantees locally optimal coordination and ensures safe operation of large and heterogeneous AGVs. Building on this concept, the proposed framework integrates a modified version of the Bounded Horizon Conflict Based Search (CBS) technique within a Rolling Horizon Conflict Resolution strategy, utilizing an extended time horizon for each agent to enable effective conflict resolution in corridors identified by a topological map. In contrast to state-of-the-art methods for AGV fleet traffic management, the proposed solution is designed for real-world, non-standardized (i.e., non-grid-like) industrial settings characterized by narrow bidirectional corridors and high-traffic density, where AGVs of various sizes and capabilities operate simultaneously. Key contributions include an anytime conflict resolution strategy with adaptive time horizon regulation, an execution layer for safe and standard-compliant interaction with real AGVs, and an advanced mechanism for deadlock detection and resolution. Experimental results obtained in realistic industrial environments demonstrate higher throughput, with improvements of up to 11% over a conventional rule-based traffic management system, a state-of-the-art industrial method, and a priority-based L-MAPF variant, while maintaining continuous operation and improved efficiency.

Visual localization in large-scale UAV scenarios is a critical capability for autonomous systems, yet it remains challenging due to geometric complexity and environmental variations. While 3D Gaussian Splatting (3DGS) has emerged as a promising scene representation, existing 3DGS-based visual localization methods struggle with robust pose initialization and sensitivity to rendering artifacts in large-scale settings. To address these limitations, we propose LSGS-Loc, a novel visual localization pipeline tailored for large-scale 3DGS scenes. Specifically, we introduce a scale-aware pose initialization strategy that combines scene-agnostic relative pose estimation with explicit 3DGS scale constraints, enabling geometrically grounded localization without scene-specific training. Furthermore, in the pose refinement, to mitigate the impact of reconstruction artifacts such as blur and floaters, we develop a Laplacian-based reliability masking mechanism that guides photometric refinement toward high-quality regions. Extensive experiments on large-scale UAV benchmarks demonstrate that our method achieves state-of-the-art accuracy and robustness for unordered image queries, significantly outperforming existing 3DGS-based approaches. Code is available at: https://github.com/xzhang-z/LSGS-Loc

Open-Vocabulary Object Navigation (OVON) requires an embodied agent to locate a language-specified target in unknown environments. Many zero-shot methods rely on frontier-candidate reasoning under incomplete observations, while topology-aware methods reduce candidate redundancy but may still introduce panoramic inspection overhead and repeated reconsideration. We present DRIVE-Nav, a structured framework that organizes exploration around persistent directions rather than raw frontiers. By inspecting encountered directions more completely and restricting subsequent decisions to still-relevant directions within a forward 240-degree view range, DRIVE-Nav reduces redundant revisits and improves path efficiency. The framework extracts and tracks directional candidates from weighted Fast Marching Method (FMM) paths, maintains representative views for semantic inspection, and combines vision-language-guided prompt enrichment with cross-frame verification to improve grounding reliability. Experiments on HM3D-OVON, HM3Dv1, HM3Dv2, and MP3D demonstrate strong overall performance and consistent efficiency gains. On HM3D-OVON, DRIVE-Nav achieves 50.2% SR and 32.6% SPL, improving the previous best method by 1.9% SR and 5.6% SPL. It also delivers the best SPL on HM3Dv1, HM3Dv2, and MP3D and transfers to a physical humanoid robot. Real-world deployment also demonstrates its effectiveness.

Stable traversal over geometrically complex terrain increasingly requires exteroceptive perception, yet prior perceptive humanoid locomotion methods often remain tied to explicit geometric abstractions, either by mediating control through robot-centric 2.5D terrain representations or by shaping depth learning with auxiliary geometry-related targets. While effective, these approaches introduce additional map-construction procedures or multi-stage skill-transfer processes beyond direct depth-to-control learning. We propose CReF (Cross-modal and Recurrent Fusion), a single-stage depth-conditioned humanoid locomotion framework that learns locomotion-relevant features directly from raw forward-facing depth without explicit geometric intermediates. CReF couples proprioception and depth tokens through proprioception-queried cross-modal attention, fuses the resulting representation with a gated residual fusion block, and performs temporal integration with a Gated Recurrent Unit (GRU) regulated by a highway-style output gate for state-dependent blending of recurrent and feedforward features. To further improve terrain interaction, we introduce a terrain-aware foothold placement reward that extracts supportable foothold candidates from foot-end point-cloud samples and rewards touchdown locations that lie close to the nearest supportable candidate. Experiments in simulation and on a physical humanoid demonstrate robust traversal over diverse terrains and effective zero-shot transfer to real-world scenes containing handrails, hollow pallet assemblies, severe reflective interference, and visually cluttered outdoor surroundings.

Robot demonstration datasets used to train vision-language-action policies can contain a subtle but harmful failure mode: trajectories that are behaviorally correct but paired with the wrong language instruction. We study post-hoc auditing of these Instruction-Trajectory Mismatches (ITMs). Unlike failed rollouts, ITMs often look plausible, and can corrupt the language-behavior mapping learned by the policy. We propose Multimodal Probabilistic Fusion (MMPF), a training-free auditing framework that treats each modality as an expert, estimates a task-label distribution from local neighborhood agreement and global prototype similarity, and then fuses modalities with predictive-entropy weighting in a product of experts. Across LIBERO benchmarks with injected instruction mismatches and noisy real-robot data, MMPF achieves the strongest overall ITM detection and label correction accuracy. We also show that auditing improves most downstream policy learning in settings where language is needed to disambiguate the task. We demonstrate in real robot experiments that our method can achieve improved policy performance and show the trade-off of filtering demonstrations compared to relabeling.

Long-horizon, high-dynamic motion tracking on humanoids remains brittle: retargeted reference motions are typically kinematically plausible but dynamically inconsistent with the robot, so small tracking errors accumulate and eventually destabilize control. We present RobotDancing, a practical single-stage reinforcement learning recipe that tracks retargeted motions by predicting reference-conditioned residual joint targets. By parameterizing actions as residual corrections on top of the reference, the policy can focus on compensating reference--robot dynamics mismatch (e.g., actuation limits, latency, friction, inertia) rather than re-synthesizing the motion. We train one policy per reference sequence while reusing a common training and deployment recipe across motions and platforms. We evaluate RobotDancing on eight LAFAN1 dance motions on Unitree G1 and conduct cross-platform experiments on H1 and H1-2. The resulting G1 policies execute long-horizon, high-energy behaviors on hardware without test-time modifications, and qualitative H1/H1-2 clips further illustrate cross-platform feasibility.

Recent developments in planetary exploration have shown the potential of Unmanned Aerial Vehicles (UAVs), such as the Ingenuity helicopter that provided valuable mapping data. However, limited payload capabilities constrain the flight times and compute available for localization, which restrict their applicability. By providing a tethered connection, issues such as battery and computational constraints are offloaded to the base rover. At the same time, the cable can be exploited for non-drifting localization. This work presents a novel Tether-Inertial Localization approach that uses tether length and angle measurements to estimate the UAV position relative to its base. The method combines a computationally efficient analytical catenary model with a Gaussian Process (GP) residual error compensation. This accounts for systematic sensor inaccuracies and model limitations. Experimental validation across circular, triangular, and figure-eight trajectories with tether lengths up to 4.5 m and a total flight time of 37 minutes demonstrates the effectiveness of the proposed approach. Using only tether-based position estimates for feedback, the analytical catenary model achieves an average RMSE of 7.4 cm, which is further reduced to 5.2 cm through GP-based residual compensation, one order of magnitude better than the state-of-the-art. These results establish Tether-Inertial Localization as a practical alternative to vision- and GNSS-based localization for Tethered Unmanned Aerial Vehicles (TUAVs).

Imitation learning from human demonstrations has achieved significant success in robotic control, yet most visuomotor policies still condition on single-step observations or short-context histories, making them struggle with non-Markovian tasks that require long-term memory. Simply enlarging the context window incurs substantial computational and memory costs and encourages overfitting to spurious correlations, leading to catastrophic failures under distribution shift and violating real-time constraints in robotic systems. By contrast, humans can compress important past experiences into long-term memories and exploit them to solve tasks throughout their lifetime. In this paper, we propose VPWEM, a non-Markovian visuomotor policy equipped with working and episodic memories. VPWEM retains a sliding window of recent observation embeddings as short-term working memory, and introduces a Transformer-based contextual memory compressor that recursively converts out-of-window observations into a fixed number of episodic memory embeddings. The compressor uses self-attention over a cache of past summary embeddings and cross-attention over a cache of historical observations, and is trained jointly with the policy. We instantiate VPWEM on diffusion policies to exploit both short-term and episode-wide information for action generation with nearly constant memory and computation per step. Experiments demonstrate that VPWEM outperforms state-of-the-art baselines including diffusion policies and vision-language-action (VLA) models by more than 20\% on the memory-intensive manipulation tasks in MIKASA and achieves an average 5\% improvement on the mobile manipulation benchmark MoMaRT. Code is available at https://github.com/HarryLui98/code_vpwem.

We present LatentAM, an online 3D Gaussian Splatting (3DGS) mapping framework that builds scalable latent feature maps from streaming RGB-D observations for open-vocabulary robotic perception. Instead of distilling high-dimensional Vision-Language Model (VLM) embeddings using model-specific decoders, LatentAM proposes an online dictionary learning approach that is both model-agnostic and pretraining-free, enabling plug-and-play integration with different VLMs at test time. Specifically, our approach associates each Gaussian primitive with a compact query vector that can be converted into approximate VLM embeddings using an attention mechanism with a learnable dictionary. The dictionary is initialized efficiently from streaming observations and optimized online to adapt to evolving scene semantics under trust-region regularization. To scale to long trajectories and large environments, we further propose an efficient map management strategy based on voxel hashing, where optimization is restricted to an active local map on the GPU, while the global map is stored and indexed on the CPU to maintain bounded GPU memory usage. Experiments on public benchmarks and a large-scale custom dataset demonstrate that LatentAM attains significantly better feature reconstruction fidelity compared to state-of-the-art methods, while achieving near-real-time speed (12-35 FPS) on the evaluated datasets. Our project page, including the code, is available at: https://junwoonlee.github.io/projects/LatentAM

Instance Image-Goal Navigation (IIN) requires autonomous agents to identify and navigate to a target object or location depicted in a reference image captured from any viewpoint. While recent methods leverage powerful novel view synthesis (NVS) techniques, such as three-dimensional Gaussian splatting (3DGS), they typically rely on randomly sampling multiple viewpoints or trajectories to ensure comprehensive coverage of discriminative visual cues. This approach, however, creates significant redundancy through overlapping image samples and lacks principled view selection, substantially increasing both rendering and comparison overhead. In this paper, we introduce a novel IIN framework with a hierarchical scoring paradigm that estimates optimal viewpoints for target matching. Our approach integrates cross-level semantic scoring, utilizing CLIP-derived relevancy fields to identify regions with high semantic similarity to the target object class, with fine-grained local geometric scoring that performs precise pose estimation within promising regions. Extensive evaluations demonstrate that our method achieves state-of-the-art performance on simulated IIN benchmarks and real-world applicability.

Generating safe corridors is essential for collision-free robotic motion planning, yet most existing methods rely on predefined reference paths, which bias corridor geometry and implicitly limit the homotopy classes that can be explored. We propose a reference-path-free corridor generation framework on graphs of convex sets (GCS) that constructs corridors directly as sequences of convex sets, allowing corridor structure to emerge from the free-space representation rather than from a guiding path. To reason about similarity among corridors, we extend visibility-based deformation from paths to convex-set sequences, enabling the fusion of topologically redundant corridors while preserving distinct alternatives. To overcome the limited adaptability of existing GCS methods based on static global decompositions, we further develop an adaptive multi-scale GCS, in which a sampling-based fine-scale graph supports localized updates and a visibility-based coarse-scale graph enables compact global exploration. The two levels maintain topological consistency, allowing incremental updates without full graph reconstruction under environmental uncertainty. Numerical experiments characterize GCS construction, corridor generation, homotopy-aware exploration, and local updates, showing efficient graph construction, stable trajectory-level performance, and shorter-duration homotopy-aware trajectories than existing baselines. Hardware experiments on ground and aerial robots, including deployment with onboard localization, further validate the framework under translated and previously unknown obstacles.

Humanoid and quadrupedal robots have the potential to revolutionize the way we work, interact, and coexist with intelligent machines. To understand their effects on society and how they can enable scientific discovery, we assess the current capabilities of these systems along hardware, locomotion, autonomy, data, and applications. We identify recent advances and key open challenges that must be overcome to enable widespread adoption and new use cases for legged robots. Last, we provide an outlook on the future of legged robots, exploring their ethical considerations, economic potential, policy implications, and broader societal effects.

Dual‐arm mobile manipulators can transport and manipulate large objects with simple end‐effectors. Interacting with dynamic environments subject to strict safety and compliance requirements, achieving whole‐body motion planning online while meeting various hard constraints for highly redundant mobile manipulators poses a significant challenge. We tackle this challenge by presenting an efficient whole‐body motion‐planning approach based on model‐based predictive control (MPC). We construct a hierarchical MPC framework in which the first MPC optimizes the end‐effectors' collision‐free motion, thereby guiding the second MPC to optimize whole‐body joint trajectories. In the first stage, we employ a Bézier‐curve representation to parameterize the high‐degree‐of‐freedom (high‐DOF) motion in of two collaborating end‐effectors, notably introducing a novel and efficient method to optimize quaternion trajectories. This facilitates fast long‐horizon motion planning of coupled translational and rotational trajectories while accounting for approximated feasibility constraints. In the second stage, this is the first work to novelly incorporate the representation into a whole‐body MPC for online high‐DOF motion generation with predictive admittance control over a relatively short horizon while satisfying whole‐body hard constraints. Compared with the usage of a discretization approach, our whole‐body MPC ensures accurate model‐state transitions with faster, more stable computation speeds, and consistent motion command generation which enhances tracking performance for our hybrid position/velocity‐controlled robot. Both MPCs perform replanning in each control loop to adapt to a changing environment. Simulations and real‐world experiments validate the proposed hierarchical MPC framework, demonstrating that its two novel MPCs achieve efficient and robust performance in scenarios involving static and dynamic obstacle avoidance, compliant interaction with manipulated objects, and external disturbances.

Navigation within complex pipelines in sectors such as construction, mining, and underwater exploration presents significant demands on robotic adaptability and maneuverability. While traditional soft robots achieve such navigation through large material deformations, they face an inherent conflict between deformation capacity and structural rigidity, which hinders the simultaneous attainment of high traversability and strong load‐bearing capacity. To address this limitation, we propose a novel structural framework to overcome existing constraints of pipeline robots through integrating the characteristics of auxetic metamaterials. Firstly, the proposed robot is monolithically fabricated via 3D printing techniques. The mechanical properties of the printing material are characterized through quasi‐static tensile tests. Subsequently, the deformation behavior of the robotic body is analyzed and compared between numerical and experimental methods. Ultimately, the robot's navigation through complex pipelines is simulated and compared with physical pipe‐passing tests. The results demonstrate that during navigation in complex pipelines, the auxetic structure of the robot deforms radially, which releases circumferential pressures. This pressure release thereby reduces motion resistance, enhancing the robot's overall maneuverability. The findings of this work transcend the performance limitations of rigid materials, thereby offering a viable pathway for robotic navigation within the complex piping networks.

This work identifies a passive aerodynamic damping effect induced by co-contraction in antagonistic redundant propulsion. Complementing prior work on aerodynamic promptness, which addressed active wrench-rate authority along constant-wrench fibers, we study the passive side: the local derivative of aerodynamic force with respect to air-relative velocity at a trim. This derivative defines an incremental aerodynamic damping coefficient. We prove that it increases monotonically along constant-force fibers under a mild aerodynamic hardening condition, and derive this property from a first-order Blade Element Theory model exposing the relevant speed-inflow coupling. The resulting mechanism, Variable Aerodynamic Damping Actuation (VADA), is formulated as an antagonistic aerodynamic actuation module and allocation principle, structurally analogous to variable-stiffness actuation at the level of fiber motions and incremental impedance modulation. An impedance-form interpretation clarifies common- and differential-mode roles, while a propeller-data-based assessment using the UIUC Propeller Database shows that the identified damping has practical small-UAV magnitude, is comparable to ordinary low-speed body-drag damping, and depends strongly on low-advance-ratio thrust sensitivity.

Salps are marine animals consisting of chains of jellyfish-like units. Their efficient underwater locomotion by coordinating multi-jet propulsion has aroused great interest in robotics. This paper presents a geometric mechanics framework for salp-inspired robots. We study a new type of geometric mechanics models inspired by salps, in which control inputs are not restricted to the shape axes, analyze nonlinear controllability, and develop motion planning and feedback control methods. We introduce the "LandSalp" robot, which serves as a physical realization of the reduced-order, drag-dominated model of salp swimming, enabling controlled evaluation of locomotion strategies without many confounding factors of underwater experiments. We extend least-squares- and inverse-dynamics-based system identification to learn the Riemannian metric of the drag-dominated model from experimental data using Lie group differentiation. With about three minutes of data, we identify an accurate model of LandSalp. Simulation and hardware experiments demonstrate omnidirectional locomotion, shape regulation, and bending maneuvers, providing a principled pathway toward more capable salp-inspired robots.

End-to-end (E2E) autonomous driving systems offer a promising alternative to traditional modular pipelines by reducing information loss and error accumulation, with significant potential to enhance both mobility and safety. However, most existing E2E approaches directly generate plans based on dense bird's-eye view (BEV) grid features, leading to inefficiency and limited planning awareness. To address these limitations, we propose iterative Proposal-centric autonomous driving (iPad), a novel framework that places proposals - a set of candidate future plans - at the center of feature extraction and auxiliary tasks. Central to iPad is ProFormer, a BEV encoder that iteratively refines proposals and their associated features through proposal-anchored attention, effectively fusing multi-view image data. Additionally, we introduce two lightweight, proposal-centric auxiliary tasks - mapping and prediction - that improve planning quality with minimal computational overhead. Extensive experiments on the NAVSIM and CARLA Bench2Drive benchmarks demonstrate that iPad achieves state-of-the-art performance while being significantly more efficient than prior leading methods.

Real-time reconstruction of deformable surgical scenes is vital for advancing robotic surgery, improving intraoperative guidance, and enabling automation. Recent methods achieve dense reconstructions from da Vinci robotic surgery videos, with Gaussian Splatting (GS) offering real-time performance via graphics acceleration. However, reconstruction quality in occluded regions remains limited, and depth accuracy has not been fully assessed, as benchmarks like EndoNeRF and StereoMIS lack 3D ground truth. We propose Diff2DGS, a two-stage framework for reliable 3D reconstruction of occluded surgical scenes. First, a diffusion-based video module with temporal priors inpaints tissue occluded by instruments with high spatiotemporal consistency. Second, we adapt 2D Gaussian Splatting (2DGS) with a Learnable Deformation Model (LDM) to capture dynamic tissue deformation and anatomical geometry, and introduce adaptive depth weight to improve geometric fidelity. We further extend evaluation beyond image-quality metrics by performing quantitative depth analysis on the SCARED dataset. Diff2DGS outperforms state-of-the-art methods in both appearance and geometry, reaching 38.02 dB PSNR on EndoNeRF and 34.40 dB on StereoMIS. Our experiments also show that optimizing image quality alone does not necessarily ensure accurate 3D geometry.The code is available at https://diff2dgs.github.io/.

Odontomas, classified as hamartomas rather than true neoplasms, are the most common type of odontogenic tumors. These lesions are typically asymptomatic, show no gender predilection, and are often detected incidentally through routine radiological examinations. Complete surgical resection is the gold standard for odontoma treatment. Traditional approaches are limited by surgeon‐dependent precision, while digital guide plate‐assisted surgery has fixed procedural constraints. This study reports the first clinical application of a semi‐active oral robotic surgical system for the minimally invasive resection of a compound odontoma embedded in the alveolar bone of the left mandibular body between the mandibular left canine (tooth 33) and first premolar (tooth 34) in a 15‐year‐old female patient. Preoperative planning was completed with high‐resolution CBCT and dedicated robotic navigation software. The odontoma was completely removed without damage to adjacent teeth. The recovery was uneventful, and a 6‐month follow‐up showed excellent bone healing with no residual lesion. The use of oral robots for odontoma extraction enables precise positioning and minimally invasive removal, offering a promising approach for the treatment of these tumors.

As a critical component for robotic citrus harvesting, the mechanical design of the end‐effector determines harvesting efficiency, success rate, and fruit quality. This paper aims to enhance citrus harvesting efficiency by proposing an integrated gripper‐shear harvesting end‐effector and its control method suitable for harvesting operations in Ehime No. 38 Citrus orchards. First, the fruits were collected and their phenotypic characteristics and mechanical properties were measured. On the basis of mechanical design theory and CAE analysis methods, the flexible clamping mechanism and the sliding shear mechanism were developed. Second, a low‐cost clamping force control method combining finite element analysis and linear regression was proposed. Its validity was verified through transient dynamic analysis experiments conducted in Ansys Workbench, involving 15 clamping experiments (10 for initial testing and 5 for validation). Finally, 50 harvesting trials conducted in citrus orchards demonstrated a 96.0% success rate for stalk severing and 94.0% success rate for harvesting undamaged citrus fruits. The proposed harvesting end‐effector is easy to control, simple to maintain, low in manufacturing cost, minimally damaging to citrus, and highly successful in harvesting.

A cooperative‐navigation algorithm based on an extended Kalman filter (EKF) is proposed for Leader–Follower autonomous underwater vehicles (AUVs) to address inconsistencies in inter‐vehicle navigation coordinate systems caused by the absence of geo‐referencing signals in anchor‐free environments. The algorithm integrates inertial compensation and a dynamic ranging threshold to address challenges posed by low update rates (0.05–0.07 Hz), high latency, time‐varying hydroacoustic channels, and multipath interference in hydroacoustic communication, all of which significantly degrade navigation accuracy. To mitigate latency from low‐rate hydroacoustic links, we use time‐tag alignment (UTC‐referenced clock synchronization) to propagate the Leader state to the effective ranging time for inertial latency compensation. In response to the hydroacoustic channel variability and multipath interference, a dynamic ranging threshold based on the AUV's motion capabilities is proposed to eliminate outlier ranging data. To enhance the robustness of cooperative navigation and maintain a coordinate system, a velocity‐bounded constraint is introduced into the EKF framework, limiting the magnitude of individual position corrections. During a field test at Danjiangkou Reservoir, Henan Province, China, AUVs cooperatively mapped an underwater terrain area of at a water depth of 45–55 m. Compared with terrain mapping using a single‐AUV inertial navigation system, the proposed cooperative‐navigation method improved the agreement of the terrain mapping with a Global Navigation Satellite System (GNSS)‐aided surface reference map by 15.89% relative to the inertial navigation baseline, which supports the effectiveness of the proposed cooperative‐navigation method during the mapping mission. Furthermore, the root mean square error (RMSE) of distance measurements between Leader AUV and Follower AUVs during linear transects decreased by 59.13% and 55.21%, respectively, indicating improved Leader–Follower range consistency achieved by the proposed method. In a Weihai open‐sea campaign with a desired Leader–Follower separation of 100 m on straight‐line segments, the resulting separation RMSE was 6.29 and 4.41 m for the two Followers, respectively. Moreover, in a Danjiangkou surface test with continuous GNSS ground truth (GNSS used only as an external reference, not as filter inputs, except for mission‐start initialization where applicable), the Followers' horizontal RMSE decreased from 15.09 m/13.00 m (INS) to 9.38 m/8.71 m (cooperative).

Swarm navigation of micro-/nanorobots has been attracting extensive attention, as it is a vital technique for microrobotic applications, for example, targeted drug delivery/therapy and micromanipulation. Researchers have shown that, controlled by a global magnetic field, millions of micro-/nanorobots can assemble and then efficiently navigate to targeted locations. However, current navigation control schemes for microswarms do not consider constraints on swarm actuation, for example, motion direction and rotating angular velocity, which would result in non-optimal navigation performance and even failure. In this work, we propose an actuation-constrained control framework for microswarms that explicitly treats important swarm motion properties as hard constraints in both trajectory planning and motion control. In our framework, we derive a constrained coarse-to-fine differential dynamic programming (DDP)-based trajectory planner that can generate the optimal trajectory in obstacle environments under the constraint of microswarm rotational angular velocity to maintain stable swarm assembly. Regarding trajectory tracking, we formulate a nonholonomic model predictive control (MPC) scheme, which makes the swarm optimally track the planned trajectory while complying with constraints on motion direction and rotational angular velocity. By our framework, microswarms can navigate with stable swarm assembly and the fastest motion speed, realizing optimal navigation performances. A series of comparative simulations and experiments validate the advantages of our framework in terms of swarm stability and navigation speed. Experimental results also show that our framework can work with different environmental morphologies, for example, discrete obstacles and channels, showing high adaptability to working scenarios. Navigation with moving obstacles and fluid flow further proves the capability of our framework for dynamic environments.

The design of a hexapod is complex and requires integration between kinematic models, control systems, and sensing. Existing literature has reviewed these sub‐systems in isolation. Since 2021, there has been no review of the field despite significant advancements in soft soil, lunar traversal, and artificial intelligence. This paper addresses that gap by establishing a novel terrain‐coded locomotion layer scaffold via taxonomy creation. A subset of 58 primary studies were deeply analyzed and synthesized to quantify method choices across hexapod locomotion. Findings show that the DH‐Parameter system and geometric inverse kinematics models are most widely used across all terrain types. Jacobian‐based inverse kinematic models prove to be computationally demanding, yet more accurate. Jacobian approaches also result in multiple end‐effector positions and unnatural poses. Bio‐inspired control systems and proprioceptive methods show promise for smoother gait switching and real‐time adaptability. Approaches related to reinforcement learning, convolutional neural networks and long short‐term memory models were also introduced in recent years. These contribute toward increasing performance, terrain adaptability and path planning. The primary contribution of this paper is a set of data‐driven design patterns linking mathematical models, layered controls, and simulator‐terrain couplings for structured and unstructured environments. These patterns support more defensible deployment‐oriented design choices. Future directions and generalized guidelines based on these design patterns for hexapod locomotion are also highlighted. Thus documenting future‐facing research which shows the hexapod becoming a fully autonomous system capable of harsh terrain traversal (lunar surfaces, ice, soft soil, underwater).

The three-dimensional divergent component of motion (3D-DCM) framework has been successfully utilized to generate center of mass (CoM) trajectories for various locomotion types. While the 3D-DCM encodes the CoM dynamics, it relies on the contact forces between the robot’s end effectors and the environment to realize the intended motion. In the original formulation of the 3D-DCM, the feasibility of contact forces concerning contact constraints is assumed, but a comprehensive analysis of this assumption is lacking. In this work, we address this gap by extending the 3D-DCM framework to incorporate contact constraints and dynamic stability of the system. This is achieved by encoding feasible CoM forces as geometric sets. We derive an analytical relationship that characterizes how these sets can be modulated by humanoid push recovery strategies. Building on these insights, we propose a push recovery algorithm that integrates ankle, hip, height-variation, and stepping strategies. The proposed method is evaluated through extensive experiments with the humanoid robot TORO, including scenarios of force-disturbed balancing, walking, and multi-contact configurations.

Non-convex constrained optimizations are ubiquitous in robotic applications such as multi-agent navigation, UAV trajectory optimization, and soft robot simulation. As a common feature in these problems, the associated non-convex constraints, including collision constraints, inversion-free constraints, and strain limits, are also non-smooth with ill-defined gradients. It is well-known that such constraints are notoriously difficult to handle, for which off-the-shelf optimizers can fail catastrophically. Instead, prior works tend to design problem-specific optimizers that trade performance for robustness. To efficiently solve this problem class in a unified manner, we propose a variant of alternating direction method of multiplier (ADMM), called BC-ADMM. Over the past decade, ADMM has achieved great success in efficiently solving many large-scale (constrained) optimization problems by decoupling them into subproblems that can be solved in parallel. However, prior ADMM algorithms lack a convergence guarantee when handling a large number of non-convex constraints with loopy constraint graphs. Instead, our BC-ADMM relaxes each non-convex constraint into a bi-convex function, further breaking the constraint into two subproblems. We show that such relaxation leads to a variant of ADMM with convergence speed guarantees under appropriate parameter choices. We further provide a practical algorithm under much milder assumptions on the parameter choices, with convergence guarantees without a speed bound. Through numerical experiments in a row of four robotic applications, we show that BC-ADMM has faster convergence than conventional gradient descent and Newton’s method in terms of wall clock time.

We present an online model-based reinforcement learning algorithm suitable for controlling complex robotic systems directly in the real world. Unlike prevailing sim-to-real pipelines that rely on extensive offline simulation and model-free policy optimization, our method builds a dynamics model from real-time interaction data and performs policy updates guided by the learned dynamics model. This efficient model-based reinforcement learning scheme significantly reduces the number of samples to train control policies, enabling direct training on real-world rollout data. This significantly reduces the influence of bias in the simulated data, and facilitates the search for high-performance control policies. We adopt online optimization analysis to derive sublinear regret bounds under stochastic online optimization assumptions, providing formal guarantees on performance improvement as more interaction data are collected. Experimental evaluations were performed on a hydraulic excavator arm and a soft robot arm, where the algorithm demonstrates strong sample efficiency compared to model-free reinforcement learning methods, reaching comparable performance within hours. Robust adaptation to shifting dynamics was also observed when the payload condition was randomized. Our approach paves the way toward efficient and reliable on-robot learning for a broad class of challenging control tasks.

Ultra-wideband radar identifies components of a muscle’s contractile state to estimate muscle force.

The distinction is a matter of life and death to both a predator and a robot in the 2025 movie Predator: Badlands .

Legged robots must achieve both robust locomotion and energy efficiency to be practical in real-world environments. Yet controllers trained in simulation often fail to transfer reliably, and most existing approaches neglect actuator-specific energy losses or depend on complex, hand-tuned reward formulations. We propose a framework that integrates sim-to-real reinforcement learning with a physics-grounded energy model for permanent magnet synchronous motors. The framework requires a minimal parameter set to capture the simulation–reality gap and employs a compact four-term reward with a first-principle-based energetic loss formulation that balances electrical and mechanical dissipation. We evaluate and validate the approach through a bottom-up dynamic parameter identification study, spanning actuators, full-robot in-air trajectories and on-ground locomotion. The framework is tested on three primary platforms and deployed on 10 additional robots, demonstrating reliable policy transfer without randomization of dynamic parameters. Our method improves the energetic efficiency over state-of-the-art methods, achieving a 32% reduction in the full Cost of Transport of anymal (1.27). All code, models, and datasets are publicly available.

Learning-based methods excel at robot motion generation but remain limited in contact-rich physical interaction. Impedance control provides stable and safe contact behavior but requires task-specific tuning of stiffness and damping parameters. We present Diffusion-Based Impedance Learning, a framework that bridges these paradigms by combining generative modeling with energy-consistent impedance control. A Transformer-based Diffusion Model, conditioned via cross-attention on measured external wrenches, reconstructs simulated Zero-Force Trajectories (sZFTs) that represent contact-consistent equilibrium behavior. A SLERP-based quaternion noise scheduler preserves geometric consistency for rotations on the unit sphere. The reconstructed sZFT is used by an energy-based estimator to adapt impedance online through directional stiffness and damping modulation. Trained on parkour and robot-assisted therapy demonstrations collected via Apple Vision Pro teleoperation, the model achieves sub-millimeter positional and sub-degree rotational accuracy using only tens of thousands of samples. Deployed in real-time torque control on a KUKA LBR iiwa, the approach enables smooth obstacle traversal and generalizes to unseen tasks, achieving 100% success in multi-geometry peg-in-hole insertion. The code for all experiments is publicly available on GitHub and videos of the experiments are available on the project website.

Quantifying the uncertainty of an object's pose estimate is essential for robust control and planning. Although pose estimation is a well-studied robotics problem, attaching statistically rigorous uncertainty is not well understood without strict distributional assumptions. We develop distribution-free pose uncertainty bounds about a given pose estimate in the monocular setting. Our pose uncertainty only requires high probability noise bounds on pixel detections of 2D semantic keypoints on a known object. This noise model induces an implicit, non-convex set of pose uncertainty constraints. Our key contribution is SLUE (S-Lemma Uncertainty Estimation), a convex program to reduce this set to a single ellipsoidal uncertainty bound that is guaranteed to contain the true object pose with high probability. SLUE solves a relaxation of the minimum volume bounding ellipsoid problem inspired by the celebrated S-lemma. It requires no initial guess of the bound's shape or size and is guaranteed to contain the true object pose with high probability. For tighter uncertainty bounds at the same confidence, we extend SLUE to a sum-of-squares relaxation hierarchy which is guaranteed to converge to the minimum volume ellipsoidal uncertainty bound for a given set of keypoint constraints. We show this pose uncertainty bound can easily be projected to independent translation and axis-angle orientation bounds. We evaluate SLUE on two pose estimation datasets and a real-world drone tracking scenario. Compared to prior work, SLUE generates substantially smaller translation bounds and competitive orientation bounds. We release code at https://github.com/MIT-SPARK/PoseUncertaintySets.

To survive in dynamic and unstructured environments, animals have developed extraordinary multimodal locomotion capabilities. This biological insight drives the innovation of bioinspired multimodal robots, which are defined as robotic systems integrating and transitioning between two or more distinct modes of bioinspired locomotion. This Review explores the historical progression, key design considerations, and current challenges faced in the realm of multimodal robotics. We highlight recent advancements in robotic body design, including the integration of soft materials, innovative structure repurposing strategies, and the deployment of multirobot systems. These advancements facilitate seamless mode transitions through both active and passive structural reconfigurations. In the realm of path planning and motion control, the paradigm is progressively shifting from traditional graph-based approaches and discrete controllers to learning-based frameworks. To address the current void of standardized benchmarks in this field, we propose five performance metrics—number of modes, marginal cost of modality, component repurpose percentage, transition cost, and performance improvement—that provide a quantitative framework for evaluating multimodal robots in terms of design effectiveness and operational performance. Last, we delineate a strategic roadmap for the future that advocates for the integration of physical and computational intelligence in multimodal robotics, arguing that this convergence is crucial for enabling real-time behavioral adaptations to complex environmental stimuli and thereby enhancing the robustness and functionality of multimodal robotic systems.

Drone light shows (DLShows) represent a rapidly growing application of swarm robotics, creating captivating aerial displays through the synchronized flight of hundreds or thousands of unmanned aerial vehicles (UAVs) as environmentally friendly and reusable alternatives to traditional pyrotechnics. This domain presents unique challenges in optimally assigning drones to visual waypoints and generating smooth, collision‐free trajectories at a very large scale. This article introduces the Unified Assignment and Trajectory Generation (UATG) framework. The proposed approach concurrently solves two core problems: the optimal assignment of drones to designated goal locations and the generation of dynamically feasible, collision‐free, time‐parameterized trajectories. The UATG framework is specifically designed for DLShows, ensuring minimal transition times between formations and guaranteeing inter‐drone collision avoidance. A key innovation is its exceptional computational efficiency, enabling the coordination of large‐scale in real‐time; for instance, it computes the optimal assignment and trajectories for 1008 drones in approximately 1 s on a standard laptop. Extensive simulations in realistic environments validate the framework's performance, demonstrating its capability to orchestrate complex formations—from alphanumeric characters to intricate 3D shapes—with precision and visual smoothness. This work provides a critical advancement for the DLShow industry, offering a practical and scalable solution for generating complex aerial choreography and establishing a valuable benchmark for ground control station software designed for the efficient coordination of multiple UAVs. A supplemental animated simulation of this work is available at https://youtu.be/-Fjrhw03594 .

During coal mining, coal and gangue with varying shapes, sizes, and masses are transporte together on conveyor belts, posing significant challenges for gangue sorting. Traditional manual sorting is increasingly limited by safety risks, labor costs, and the shortage of skilled workers. To address this problem, this paper presents a push‐based gangue‐sorting robotic system for industrial dynamic conveyor scenarios. The system is organized as a task‐specific closed‐loop robotic sorting framework that couples gangue perception, identity‐consistent tracking, contour‐level pose estimation, time‐aware push planning, coal‐carryover avoidance, and multimodal execution‐outcome verification for dynamic mine‐conveyor operation. In contrast to conventional grasp‐lift‐place sorting systems, the proposed system exploits conveyor support and uses planar pushing as the manipulation primitive, which is better suited to large, heavy, and irregular gangue pieces in constrained mine environments. Long‐term tests show that the proposed system can stably complete gangue separation under isolated deployment conditions without manual intervention. In industrial online tests, the recognition accuracy for dynamic gangue exceeds 92%, the average planning time of a single robotic arm is about 0.4 s, the average execution time is about 2.8 s, and the highest single‐arm gangue removal rate reaches 61.52%. A two‐stage cascade estimate based on the highest single‐arm result suggests the potential for line‐level removal above 80%. The proposed execution‐outcome verification module achieves verification accuracy exceeding 95%.

Agricultural robots are increasingly used to improve orchard productivity and reduce labor dependence, yet collaborative orchard mapping remains difficult because dense canopy degrades GNSS reliability, repetitive row geometry induces structural aliasing, and long‐duration operation over large orchard blocks causes mergeable overlap to emerge only after extended traversal, thereby destabilizing generic registration pipelines. In this work, we formulate orchard map merging as an operation‐time progressive submap association and registration problem for a dual‐robot field setting and propose PROMO‐Orchard, an orchard‐aware semantic map‐merging framework. The framework integrates four tightly coupled components: class‐dependent voxelization with persistent orchard‐prior extraction, Orchard‐AKS for operation‐aware submap scheduling, Sem‐OREOS for semantic‐enhanced overlap retrieval and merge triggering, and Sem‐GICP for class‐weighted semantic refinement with orchard structural regularization. Together, these components turn semantic map merging into an operation‐time, orchard‐aware process rather than a generic post‐hoc registration step. Field experiments in three representative orchard settings, namely I‐Tunnel, U‐Tunnel, and Long‐Duration, show that the proposed method consistently suppresses row aliasing and maintains semantically coherent shared maps under in‐row overlap, headland U‐turning, and delayed‐overlap operation. Compared with Segregator and a semantic‐ablation variant, the proposed framework reduces ICP error by 78–94% while maintaining FPMR above 91% across all scenarios, and achieves clear improvements in semantic fidelity, boundary continuity, and structural consistency. These results support the conclusion that reliable orchard map merging requires not only semantic cues, but their structured integration with orchard‐specific priors, thereby providing a practical basis for collaborative orchard mapping and downstream robotic operations such as navigation, canopy management, harvesting, and logistics. While the present study experimentally validates the dual‐robot case, the proposed progressive pairwise merging formulation provides the system‐level basis for future extension to larger multi‐robot orchard teams.

Efficient material logistics are critical in controlling costs and schedules in the construction industry. However, manual material handling remains prone to inefficiencies, delays, and safety risks. Autonomous forklifts offer a solution to streamline on‐site logistics, reducing reliance on human operators and mitigating labor shortages. This paper presents the development and evaluation of ADAPT (Autonomous Dynamic All‐terrain Pallet Transporter), an autonomous off‐road forklift designed for construction environments. Unlike structured warehouses, construction sites pose significant challenges, including dynamic obstacles, unstructured terrain, and varying weather conditions. To address these challenges, our system integrates AI‐driven perception with traditional approaches for decision making, planning, and control, enabling reliable operation in complex environments. We validate the system through extensive real‐world testing, comparing its performance against an experienced human operator across various weather conditions. Our findings demonstrate that autonomous outdoor forklifts can operate near human‐level performance, offering a viable path toward safer and more efficient construction logistics.

Degeneracy caused by the repetition of geometric features remains a significant bottleneck in robotic state estimation systems like odometry and SLAM. To address this challenge, we propose a novel approach based on screw theory to model feature constraints and analyze degeneracy. Unlike conventional methods, our approach derives feature constraint representations via geometric operations rather than by derivative computations, and decouples the cost-function formulation, yielding a more robust and interpretable framework. Initially, a feature constraint model is formulated using the screw representation. Subsequently, a degeneracy analysis model and a judgment formula are presented, both grounded in screw theory and the feature constraint model. The proposed model inherently separates translational and helical degeneracy while accurately estimating environmental parameters, such as the position and pitch of the rotation axis, to reduce assessment errors, enable outlier rejection, and improve the robustness of degeneracy detection. A novel feature extraction and management method is proposed to improve computation efficiency and coverage of the feature in sparsely scanned scenarios. Simulations and real-world experiments show that our approach improves the robustness and accuracy of degeneracy detection while reducing the computation time by more than 70%. A dimensionless stability metric for the judgment formula is proposed, showing that our approach significantly outperforms existing state-of-the-art (SOTA) approaches by more than two orders of magnitude. To date, this study constitutes the first formal endeavor to conceptualize and systematically analyze the problem of helical degeneracy, providing a novel and rigorous perspective on the inherent challenges of robotic state estimation.

COMFI (human–robot Collaboration Oriented Markerless For Industry) is a multimodal dataset designed to advance markerless motion capture, ergonomics, and Human–Robot Collaboration (HRC) in factory settings. COMFI contains 5.4 h of synchronized and spatially co-registered streams acquired from 18 participants performing 24 tasks that span everyday movements (e.g., walking, sit-to-stand) and ergonomically demanding industrial tasks (lifting, overhead work, screwdriving, polishing, welding), with the addition of two HRC scenarios in which a Franka Emika Panda is guided by the human while holding a tool. For a total of 86.5 GB of data, it includes: calibrated multi-view RGB videos (40 Hz), optical motion capture markers and joint centers positions, as well as joint angles (100 and 40 Hz), 6D ground reaction forces (1000 and 40 Hz), and robot telemetry (200 and 40 Hz). Camera intrinsics/extrinsics, global triggers, and software-barrier synchronization for webcams are distributed, along with participant-scaled human Universal Robot Description Files that adhere to International Society of Biomechanics conventions, enabling kinematics, dynamics, and torque estimation. Videos are anonymized while preserving facial cues useful to markerless pipelines. Accompanying code supports loading, calibration, and visualization. COMFI enables rigorous benchmarking of markerless pose estimation under occlusion and clutter against reference systems, allowing the extension of current state-of-the-art algorithms to complex industrial scenarios. COMFI is expected to catalyze reproducible, cross-disciplinary research toward safer, more ergonomic HRC. For a more illustrative summary, go to https://comfi-gepetto.github.io/ .

The human pose estimation technology based on robotic vision sensing systems has emerged as a critical perception approach for collision avoidance and safety assurance in human‐robot collaboration (HRC). In dynamic occlusion scenarios, human pose estimation primarily relies on rapidly evolving deep learning models, yet it still faces challenges such as degraded recognition accuracy under prolonged occlusion and the high time cost of data collection and annotation. To address this issue, we proposed a novel skeletal pose and minimum‐distance fusion (SPMF) approach which aims to achieve robust human pose estimation in dynamically occluded human‐robot collaboration (HRC) scenarios. Firstly, the 3D coordinates of the joints were pre‐estimated using the OpenPose‐based skeletal model, in which the human body was represented by capsules at the joint positions. Following, the minimum human‐robot distance was calculated via the Gilbert‐Johnson‐Keerthi (GJK) algorithm by combining the capsule model of the robot. In addition, a framework for human postural discrimination was established by combining the joint prediction confidence and minimum distance parameters, which can judge the reliability of pre‐estimated pose data for occluded points. Finally, an algorithm for correcting misjudged joints was proposed which reconstructed the depth values for fully occluded joints. The experiments conducted on a refined Human3.6M‐Occluded‐SV database show that the proposed method substantially outperforms other fusion methods under severe occlusion. Furthermore, the experiments on real HRC scenarios demonstrate that the pose estimation framework can achieve real‐time collision detection under robot occlusion and deliver accurate pose recognition for industrial safety applications.

The cover image is based on the article TriRock6W: Autonomous Mobile Robot with Six Wheels, Three Rocker Arms in Complex Environments by shao shiliang et al., https://doi.org/10.1002/rob.70229.

In swarm robotics, stigmergy (indirect communication through changes to the environment) offers an alternative to direct communication and centralised control when these are unavailable. We investigate whether stigmergy alone can solve a group-formation task, namely controlling the number of robots that gather into a group. To this end, a leader robot lays a phosphorescent trail by exciting the arena surface with ultraviolet light; the brightness of the trail decays over time, and its initial intensity encodes the target group size. Following robots join the trail and leave it once the measured brightness falls below a threshold, so that the group size is regulated without any direct inter-robot communication. We demonstrate this mechanism as a proof of concept on real hardware, using a swarm of small Dezibot robots, and show that groups of a specified size form and are maintained over an extended period of time.

3D object detection using light detection and ranging (LiDAR) sensors requires a balance between accuracy and computational efficiency for onboard perception in autonomous driving and robotic navigation. Many existing LiDAR-based detection methods employ complex architectures to extract features, integrating large amounts of contextual information to enhance accuracy. This often results in significant computational costs, leading to suboptimal performance on resource-constrained embedded devices. In this study, we propose a knowledge distillation framework that transfers object-level voxel representations from a strong teacher model to lightweight student models through selective voxel-space feature alignment. Taking advantage of the linear-time sequence model with selective state spaces (Mamba), we design a multi-branch Mamba teacher backbone and a box-aware feature transfer mechanism that aligns spatially corresponding voxel features between teacher and student networks through a Mamba-based projection module. Experimental results on both a public dataset and real-world data show that our approach significantly reduces computational load while maintaining competitive accuracy compared with state-of-the-art methods.

Robot localization is the process of positioning a mobile robot in relation to its surroundings, and its precise positioning is essential for decision‐making and navigation. An autonomous robot's core competency is localization because the robot's positioning is significant in deciding its future decisions and navigation, but existing techniques lack accuracy in outdoor environments. To overcome this, in this manuscript, Dual‐Discriminator Conditional GAN and Optimized Binarized Spiking Neural Networks for Accurate Outdoor Robot Localization and landmark detection (DGAN‐LD‐OBNL) are proposed. This work addresses the critical need for precise positioning of robots in various applications by leveraging two novel deep learning techniques and a transfer learning technique. Here, outdoor robot localization data are collected from the virtual KITTI data set. Afterwards, Dual‐discriminator Conditional Generative Adversarial Network is proposed for precise landmark detection that significantly improves the localization accuracy. Then, Binarized Spiking Neural Network with EffiectiveNetV2 (BSNN‐ENetV2) is proposed to determine the robot location and compass direction. The weight parameters of BSNN‐ENetV2 are optimized using Volcano Eruption Optimization Algorithm for enhancing detection accuracy. The experimental outcomes show that the proposed method outperforms existing techniques in both accuracy, computational time, thus making it a promising solution for autonomous robot localization in outdoor environments.

Memory-augmented robotic policies are essential in handling memory-dependent tasks. However, existing approaches typically rely on simply extending the observation window, struggling to simultaneously achieve precise task-state tracking and robust long-horizon retention. To overcome these challenges, inspired by the Atkinson--Shiffrin memory model, we propose MemoAct, a hierarchical memory-augmented policy that leverages distinct memory tiers to tackle specific bottlenecks. Specifically, sensory memory filters immediate perceptual inputs, lossless short-term memory supports precise task-state tracking, and compressed long-term memory facilitates robust long-horizon retention. To enrich the evaluation landscape, we construct MemoryRTBench based on RoboTwin 2.0, comprising 6 manipulation tasks that systematically evaluate policy memory capabilities across three dimensions: sequential, spatial, and episodic memory. Extensive experiments across simulated and real-world scenarios demonstrate that MemoAct achieves superior performance compared to both existing Markovian baselines and history-aware policies. The project page is available at https://tlf-tlf.github.io/MemoActPage/.

Autonomous navigation in farmland is difficult because weak texture, crop occlusion, moving machinery, and illumination changes affect perception and control simultaneously. This study develops an adaptive unmanned aerial vehicle obstacle‐avoidance framework comprising a MobileViT‐XXS visual encoder, calibrated millimeter‐wave‐radar fusion, spatiotemporal trajectory prediction, an RRT* global planner, and a Soft Actor–Critic (SAC) local controller with semantic safety constraints. Under the matched‐SAC simulation protocol, the proposed perception backbone achieved an average mAP of 87.4%, exceeding YOLOv8n + SAC by 5.8 percentage points and EfficientViT‐L2 + SAC by 3.2 percentage points; the paired scene‐level difference relative to EfficientViT‐L2 + SAC was 3.2 percentage points (95% CI, 1.1–5.3; Holm‐adjusted p = 0.006). The complete simulation framework yielded an obstacle‐avoidance success rate of 88.3%, an average path length of 138.4 m, an energy consumption of 14.2 Wh/km, and a path smoothness of 0.21 rad/m. The matched simulation success‐rate difference relative to YOLOv8n + SAC was 4.8 percentage points (95% CI, −1.2 to 10.8; Holm‐adjusted p = 0.112), and therefore does not establish statistical superiority. Separately, the expanded field ledger contains 100 independent test flights across six scenario groups, below the prespecified target of at least 160; the field evidence is therefore reported descriptively. At 1920 × 1080 resolution, end‐to‐end latency was 49.2 ± 2.5 ms, whereas 3840 × 2160 required 92.1 ± 3.8 ms. A 60‐min 1080p sustained‐load test triggered thermal throttling at approximately 20 min and increased latency from 48.5 to 56.3 ms by 60 min. Accordingly, 1080p is the default closed‐loop mode and 4K is restricted to low‐speed inspection and ablation. The joint reward heatmaps and structured failure log further delimit the tested operating envelope. The results support technical feasibility under the evaluated conditions but not broad agricultural generalization or deployment readiness.

Real-world learning can push robot manipulation beyond imitation alone.

Wall‐climbing robots are increasingly used in hazardous, hard‐to‐reach environments for tasks such as crack detection, hull welding, and exterior surface cleaning. As a result, the development of wall‐climbing robots capable of robust environmental adaptation, reliable adhesion, and autonomous motion control has become an important research focus. This review first examines the environmental adaptability of wall‐climbing robots, highlighting their capabilities in wall‐to‐wall transitions, navigation across curved and irregular surfaces, and obstacle negotiation, all of which are critical to enhancing operational efficiency and reliability. Subsequently, the adhesion technologies employed by wall‐climbing robots are reviewed, with particular emphasis on negative‐pressure, magnetic, and bio‐inspired adhesion mechanisms. A comparative analysis of adhesion principles, performance metrics, and load‐bearing capabilities reported in existing studies is presented to identify suitable adhesion solutions for different operating environments and to improve stability during locomotion. Finally, critical challenges and research gaps are highlighted, and a systematic discussion is provided on the relationships between locomotion modes and adhesion mechanisms, as well as between autonomy and system maturity.

People with physical disabilities face significant mobility challenges while climbing uneven surfaces and staircases. In these circumstances, traditional wheelchairs do not provide stable, safely controlled mobility, underscoring the importance of building superior stair‐climbing wheelchairs for both medical and infrastructure purposes. In this study, a review is proposed that provides a comprehensive summary of wheelchair types, developments, compares and evaluates existing stair‐climbing wheelchair technologies, and highlights major design gaps, performance limitations, and future research directions. The review examines four fundamental categories of existing systems: tracked‐based mechanisms, wheel‐leg hybrid mechanisms, tri‐wheel mechanisms, and automatic stair‐climbing systems. These mechanisms are evaluated based on critical performance metrics, including safety and stability, ergonomics, ease of use, energy efficiency, load‐carrying capacity, cost, effectiveness, adaptability to variable terrain, stair‐climbing mobility performance, and control methodologies ranging from manual to AI‐enabled modes. The key findings indicate that although significant advancements in stair‐climbing wheelchairs have been made, current systems still face challenges in portability, weight, cost, and comfort, while recent trends highlight the growing adoption of lightweight materials, hybrid locomotion, and AI‐based navigation frameworks. The proposed review has identified research gaps, including insufficient user‐centric design, high demand for low‐cost, reliable real‐world solutions, and the need for standardized safety testing methods for stair‐climbing systems. Overall, this review will provide an overview of the different technologies and a comprehensive comparison, highlight the major design trade‐offs, and suggest future research directions to advance the design and support designers, rehabilitation engineers, policymakers, and the mobility assistance technology research community in developing user‐friendly, comfortable stair‐climbing wheelchairs.

A wide‐area seabed survey was conducted using autonomous underwater vehicles (AUVs) in a known dumpsite off the coast of Southern California that has been found to have widespread sediment contamination from the chemical dichlorodiphenyltrichloroethane (DDT). The AUVs were equipped with synthetic Aperture Sonar (SAS), an electronic still camera (ESC), and a laser line scanner to characterize fine‐scale spatial variations of the surficial sediments to gain insight into their mobility. Surveys covered of the known dumpsites to assess sediment texture and distribution patterns. In parallel, a remotely operated vehicle (ROV) performed targeted optical surveys in areas identified by the AUV. The ROV collected 300 h of high‐definition video footage, providing more detailed seafloor observations that were not captured with the ESC. This study employs machine learning and seafloor roughness characterization techniques on SAS data that are informed by the other modalities of seabed sensing. The physics‐informed techniques are applied for seafloor texture classification and provide new insights for interpreting SAS imagery and the seabed in the vicinity of these dumpsites that are of emerging environmental concern.

Underwater multi‐robot and swarm systems require advanced methodologies for controlling collective behavior. Conventional single‐robot techniques, such as tethering, sonar imaging, and acoustic localization, are not scalable or effective for swarm applications. This work introduces novel approaches for visual detection, relative localization, and control of a system of multiple autonomous underwater robots. In particular, we focus on agent‐to‐agent visual localization to inform control strategies and enable the emergence of coordinated swarm behaviors. The proposed methodologies have been implemented and tested on a fleet of autonomous underwater robots developed as part of the HSURF project. We present a thorough experimental validation of the integrated detection‐to‐control system, demonstrating its robustness and effectiveness in realistic, real‐world‐like conditions. This research marks a significant step forward in developing scalable solutions for underwater multi‐robot systems.

Accurate odometry estimation is fundamental to autonomous navigation, enabling precise localization and continuous motion tracking. While vision‐based and lidar‐based odometry systems have reached a high level of maturity, their performance deteriorates significantly in adverse environmental conditions, such as low visibility, inclement weather, or dynamic lighting. In contrast, 4D radar offers inherent robustness to such challenges, along with extended range and velocity‐sensing capabilities, making it a compelling alternative for long‐term, all‐weather perception. In this paper, we present an end‐to‐end neural network‐based odometry framework that directly leverages raw ADC data from 4D radars, surpassing conventional methods that are limited to processing only 4D point cloud information. We introduce a lossless encoding scheme that converts the raw ADC signals into structured heatmap representations, preserving essential spatial and temporal information. To extract meaningful features from this rich representation, we design a hybrid attention architecture that combines intra‐frame self‐attention for enhanced spatial understanding with inter‐frame cross‐attention to model fine‐grained temporal dependencies. Comprehensive evaluations on the Coloradar dataset show that our method significantly outperforms state‐of‐the‐art radar‐based odometry approaches based on traditional feature matching or deep learning. It achieves superior accuracy and robustness, highlighting its strong potential for real‐world deployment in autonomous navigation systems under challenging conditions. Our code is available at https://github.com/MoYuGit/Deep_Radar_Odometry .